Three-dimensional attitude prediction method based on pseudo-image sequence evolution

Abstract

The invention discloses a three-dimensional attitude prediction method based on pseudo-image sequence evolution. The method comprises the following specific steps: step 1, inputting a joint point coordinate sequence; step 2, converting the joint point coordinate sequence into an image sequence, dividing a human body, and converting the image sequence to obtain a pseudo image sequence; step 3, constructing a sequence-to-sequence model, and predicting a future attitude sequence according to the pseudo-image sequence of the historical attitude; and step 4, outputting an articulation point coordinate sequence of the future attitude sequence. The invention provides a three-dimensional attitude prediction method based on pseudo-image sequence evolution. The method has (1) a new skeleton representation, and an attitude prediction problem being modeled as a video prediction problem; and (2) a new structure PISEP2, which predicts all future frames in a non-recursive manner, can effectively avoid error accumulation and improve the calculation efficiency.

Description

technical field [0001] The invention relates to the technical field of attitude prediction, and more specifically relates to a three-dimensional attitude prediction method based on pseudo image sequence evolution. Background technique [0002] Attitude prediction is widely used in human-machine collaboration, home service robots, intelligent security and other fields. It is important to predict future dynamics before an action takes place, giving the robot more time to react and prepare in advance. [0003] In the traditional attitude prediction based on mocap data, mocap data is difficult and expensive to obtain, and requires a large number of preprocessing operations to visualize its prediction performance, which is relatively inefficient and not intuitive; the existing skeleton representation, in human motion In the process, the joint points of the same part may have a strong correlation, and the joint points of different parts may have a weak correlation. For example, ...

AI Technical Summary

Problems solved by technology

[0003] In the traditional attitude prediction based on mocap data, mocap data is difficult and expensive to obtain, and requires a large number of preprocessing operations to visualize its prediction performance. This processing method is relatively inefficient and not intuitive; In the process, the joint points of the same part may have a strong correlation, and the joint points of different parts may have a weak correlation

For example, there may be a strong correlation between the joint points of the left upper limb, and a weak correlation between the joint points between the upper limb and the lower limb; in the advanced sequence-to-sequence model, multiple frames are predicted in a recursive manner, It is prone to problems such as large amount of calculation and error accumulation. In addition, as the length of the input or output sequence increases, the complexity of the model increases significantly.

Images